-
-
Describe the newsletter you wish existed in one sentence
-
Nemotron picks where to look, you add the sites and communities you trust
-
A Nebius Serverless Job researches and writes each issue on its own
-
Every story told once, with a verified date and a link to its source
-
Under the hood, the brief and the 8-step Nemotron pipeline with real costs
-
Every Nemotron call on Token Factory, routed to the cheapest model that does the job
Inspiration
I'm a Mac user and I follow a lot of Apple newsletters. Every time Apple launches a new device, my inbox fills up with a dozen emails telling me the same news. I wanted one issue that tells each story once.
What it does
Describe the newsletter you wish existed in one sentence. On your delivery days, an agent searches the live web, checks every publication date, drops anything you've already read, and writes one issue just for you. Vote on stories and the next issue adapts.
How we built it
- NVIDIA Nemotron on Nebius Token Factory, with each step sent to a different model: Nemotron 3 Nano scores every search result, Nemotron 3.5 Lightning plans the searches and uses its 1M-token context to remember the reader's whole history, and Nemotron 3 Ultra writes the issue.
- Tavily searches news, blogs, Reddit, YouTube and the reader's trusted sites, limited to results published since the last issue.
- Nebius Serverless Jobs write each issue in the background.
- Next.js on Vercel, Postgres on Supabase.
Challenges we ran into
Search engines get publication dates wrong, so we open every promising result and read the real date on the page. Spotting "the same news, worded differently" needed reasoning turned on.
Accomplishments that we're proud of
Sending each step to the right model makes an issue cost $0.068, against $0.41 on Ultra alone. Every issue keeps a full trace of each search, result and model call.
What we learned
Turning reasoning off cuts output tokens about 100x on mechanical tasks, but it breaks judgment calls. We keep it on wherever the model has to decide something.
What's next for Newsletter for One
Email delivery, clarifying questions when a topic is too broad, and podcasts as a source once an NVIDIA speech model is available on Token Factory.
Built With
- nebius-ai-cloud
- nebius-serverless-jobs
- nebius-token-factory
- nemotron-3-nano
- nemotron-3-ultra
- nemotron-3.5-lightning
- next.js
- node.js
- nvidia-nemotron
- postgresql
- react
- supabase
- tavily
- typescript
- vercel
